InsectCV: um sistema para detecção de insetos em imagens digitais

Abstract

The manual task of counting and identifying small insects, such as aphids and parasitoids, captured in color-type field traps is exhausting, time-consuming, and non-scalable. This activity involves the separation of the elements of interest and requires the use of magnifiers or microscopes. Recent advances in artificial intelligence, image processing, and high-performance computing have enabled the development of efficient computer vision solutions to monitor pests and identify diseases in plants. With this in mind, this work presents InsectCV, a system for the automatic counting and identification of insects in images generated by the scanning of samples captured in traps. For the development of this solution, we used a 209 grayscale images dataset containing 17,908 labeled insects, a convolutional neural network Mask R-CNN to generate the model, and the development of three web services. During the training of the model, we applied the transfer learning technique and the data augmentation. We defined two new parameters to adjust the false-positive ratio by class. We used images of insects obtained from traps exposed in Coxilha and Passo Fundo, RS, Brazil in 2019 and 2020 wheat crops. In comparison to the manual counting, we verified coefficients determination close to 1 (R2 = 0:87 for aphids and R2 = 0:92 for parasitoids), proving the ability of the model to identify the fluctuation of population levels for these insects. Therefore, InsectCV can be used to detect action thresholds in alert systems for integrated pest management.